Resource management for ultra-scale computational grid applications

Abstract

Advances in networking infrastructure have made it possible to build very large scale applications whose execution spans multiple supercomputers. In such very large scale or ultra-scale applications, a central requirement is the ability to simultaneously co-allocate large collections of resources, to initiate a computation on those resources and to initialize the distributed collection of components to construct a single, integrated computation. In a previous paper [3], we defined a general resource management architecture for high-performance distributed systems in which resource co-allocation was an integral component. In this extended abstract, we examine co-allocation in more detail and describe the implementation of a specific resource co-allocator called the Dynamically Updated Request Online Co-allocator, or DUROC. DUROC has been implemented as part of the Globus grid toolkit. We briefly describe the design of DUROC and discuss how is has been used to support a range of large-scale grid applications.